xAI Grok Voice Transcribe 2.0 Doubles STT Accuracy at Unchanged Pricing

xAI released Grok Voice Transcribe 2.0, a new speech-to-text model for its Speech-to-Text API. The vendor claim that matters for builders is simple: across xAI’s real-world evaluations, 2.0 is about twice as accurate as Grok Voice Transcribe 1.0 at the same price. Batch stays $0.10 per hour of audio; streaming stays $0.20 per hour, with diarization, timestamps, and key-term biasing included.

That combination — accuracy jump without a price bump, plus a default-model swap coming soon — is why this is today’s piece. It is an API ship you can wire into call centers, voice agents, and video pipelines, not a teaser slide. Primary source: Introducing Grok Voice Transcribe 2.0.

Image credit: xAI

Key points

• What shipped: Grok Voice Transcribe 2.0 STT model; ~2× accuracy vs 1.0 at unchanged pricing (xAI).
• Foundation: Built on the audio stack behind Grok Voice — customer-support calls, video narration, Tesla Grok assistant — trained on live, noisy, multilingual real-world audio.
• Public claim: Ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard (per xAI).
• Internal evals (xAI): Improves vs 1.0 on telephony, conversational (Grok), credentials (phones/emails/addresses), and short multilingual voice commands (19 languages); telephony leads models they tested.
• Named metric: Short-phrase WER 20.6% → 6.8% vs 1.0.
• API surface: Batch (POST https://api.x.ai/v1/stt) + streaming WebSocket (wss://api.x.ai/v1/stt); word-level timestamps + confidence; speaker diarization (no extra cost); multichannel up to 8; key-term biasing up to 100; text formatting; filler removal; smart turn detection.
• Models: grok-voice-transcribe-2.0 (default when omitted); pin grok-voice-transcribe-1.0 to stay on 1.0. 1.0 deprecated soon.
• Limits: Max file 500 MB; WAV/MP3/OGG/Opus/FLAC/AAC/MP4/M4A/MKV + raw pcm/mulaw/alaw.
• Customer story: Atlassian Loom using 2.0 for video transcripts; Sanchan Saxena (SVP, Teamwork Collection) on Loom → Cursor workflows.
• Caveat: Benchmarks are largely vendor-framed; HN notes skepticism on STT leaderboards and that weights are not open. 9to6AI has not hands-on tested 2.0.

What shipped

xAI positions Grok Voice Transcribe 2.0 as a production STT model for hard audio: flaky phone lines, competing voices, accents, and spoken credentials. The model sits on the same audio foundation that already powers Grok Voice workloads — tens of thousands of customer-support calls a day, millions of hours of video narration, and in-product voice agents including Tesla’s Grok assistant (per xAI).

For integrators, the practical ship is the Speech-to-Text API:

Docs list language codes for text formatting (numbers, currencies, units into written form). The model can still transcribe those languages without the language flag; setting it turns formatting on. Auto language detection and mid-recording language switches are part of the 2.0 product story; multilingual accuracy is called out as the largest improvement over 1.0. Full reference: Speech-to-Text docs.

What changed vs 1.0

Accuracy (vendor-reported). xAI says 2.0 is roughly twice as accurate as 1.0 at the same price, and that it ranks first among 32 streaming models on Artificial Analysis’s public leaderboard. Treat the leaderboard rank as xAI’s citation of a public board, not an independent 9to6AI audit.

Internal sets drawn from production traffic (again, xAI):

1. Telephony (8 kHz) — customer-support calls, English; 2.0 leads models they tested.
2. Conversational — conversations with Grok, English.
3. Credentials — phone numbers, emails, addresses, English.
4. Short phrases — voice-assistant utterances across 19 languages.

The one hard number xAI publishes in prose for the short-phrase set: word error rate drops from 20.6% to 6.8% versus 1.0. Charts on the announcement page compare 2.0 to 1.0, ElevenLabs Scribe v2, and Deepgram Nova-3; we are not inventing extra WER figures beyond what the post states in text.

Features (same price envelope). Existing STT integrations get the accuracy bump with no code changes if they ride the default model. Controls called out:

• Batch and streaming
• Word-level timestamps with confidence scores
• Speaker diarization at no additional cost
• Multichannel transcription up to 8 channels
• Key-term biasing (up to 100 domain terms per request)
• Text formatting / inverse text normalization
• Filler word removal (um / uh off by default)
• Smart turn detection for voice agents (WebSocket smart_turn threshold + optional timeout)

Pricing unchanged. Batch $0.10/hr; streaming $0.20/hr; diarization, timestamps, and key terms included. xAI says 2.0 will soon be the default in the Speech-to-Text API and that 1.0 will be deprecated in the coming weeks — pin grok-voice-transcribe-1.0 if you need a freeze during the transition.

Customer proof point. Atlassian found 2.0 more accurate than their prior solution for Loom video transcripts. Sanchan Saxena, SVP of Teamwork Collection at Atlassian, framed the loop as: capture context once in Loom, pipe the transcript into Cursor, and close “context to code.” That is a vendor customer quote, not a 9to6AI benchmark.

What it means for builders

If you already call xAI STT, the default-model change is the operational headline. Omitting model will land on 2.0; if your eval suite or compliance story depends on 1.0 behavior, pin the slug now and schedule a re-eval before deprecation.

If you are choosing a provider, the useful comparison frame is not “another chatbot voice” — it is noisy telephony, credentials, short multilingual commands, and agent turn-taking at a published hourly rate. Included diarization and multichannel (up to 8) matter for contact-center and dual-track recordings without a second SKU. Key-term biasing (product names, medical terms) and formatting are the usual production levers that decide whether transcripts are paste-ready or need a cleanup model.

Streaming docs are builder-dense: interim results, endpointing, Opus encoding for bandwidth-constrained clients, per-channel finalize, and Smart Turn to avoid cutting users off mid-number. That is the surface for voice agents and live captions — not just offline file dumps.

Honest limits for planning:

• No open weights. The HN thread’s “no weights” note is fair: this is an API product, not a self-host drop.
• Vendor evals. Artificial Analysis rank and internal WER charts come through xAI’s post. Community STT debates (including skepticism that public boards underrate models like Nova-3) are a reminder to run your audio before locking a vendor.
• We have not hands-on tested 2.0. Numbers and feature lists above are from the official announcement and docs.

What to watch

1. Default cutover and 1.0 deprecation window — confirm in docs/changelog when the pin becomes mandatory for 1.0 behavior.
2. Independent benches — Artificial Analysis updates and third-party WER on telephony / credentials, not only clean read speech.
3. ZDR / enterprise controls — HN asked for zero data retention; not the focus of this announcement.
4. Loom / Cursor-style workflows — whether more product teams publish transcript → coding-agent loops with measurable latency and error rates.
5. Competitive price moves — HN commenters already compared $0.10/hr batch to higher Gemini Transcribe pricing; watch whether rivals match included diarization at this rate.

Primary sources

• Introducing Grok Voice Transcribe 2.0 (xAI)
• Speech to Text — xAI docs
• Hacker News discussion (~24 points; light engagement; benchmark skepticism + no-weights note)